On the Trade-offs between Adversarial Robustness and Actionable Explanations

Adversarial robustness and explainability may conflict: The paper provides evidence that making AI models more resistant to adversarial attacks (adversarial robustness) can make it harder to provide people with clear, actionable explanations of how to change a model's decision in their favor.

Recourse becomes more difficult and less reliable: When models are adversarially robust, the "recourse" (specific actions users can take to change a negative prediction to positive) becomes more costly to implement and less likely to actually work. The paper demonstrates this through both theoretical analysis and real-world dataset testing.

Trade-offs exist between two important safety goals: This research highlights a fundamental tension in responsible AI: organizations cannot easily maximize both adversarial robustness (protection against malicious attacks) and actionable explanations (transparency and user agency) simultaneously—they must make deliberate choices about which to prioritize.

Real-world implications for high-stakes decisions: In critical applications like loan approvals or hiring decisions, this trade-off matters significantly. Users denied by a robust model may receive explanations that are impractical or ineffective, limiting their ability to understand and contest decisions affecting their lives.

The problem scales with model complexity: The findings apply to both simple linear models and more complex non-linear models, suggesting this is a widespread challenge rather than a limitation of specific model types.